The U.S. Is Turning AI Export Controls Into Compute Governance

What changed

The Commerce Department’s Bureau of Industry and Security has clarified that advanced-computing export-license requirements follow the buyer’s headquarters and ultimate parent, not just the shipping address. In guidance dated May 31, 2026, BIS says a license is required to export covered advanced-computing items to entities headquartered in Country Group D:5 or Macau, or whose ultimate parent is headquartered there, even if the entity receiving the item is physically in a third country.

That matters because Reuters reported the move as an effort to block Nvidia and AMD AI-chip shipments to Chinese firms operating outside China, including through overseas subsidiaries in places such as Malaysia. The covered items include advanced-computing ECCNs such as 3A090 and 4A090 categories and related items.

This should not be described as simply a new ban. BIS frames the document as enforcement guidance for a license requirement first introduced on November 17, 2023. The policy trajectory began with the October 2022 advanced-computing and semiconductor controls, then tightened in October 2023 to address performance thresholds, loopholes, and durability.

The target is offshore compute access

The practical issue is not whether a chip crosses directly into China. The issue is whether a Chinese-headquartered firm can obtain or control frontier compute through an overseas subsidiary, cloud partner, server integrator, or data-center arrangement. A model can be trained outside China if the organization can secure enough controlled GPUs elsewhere.

That is why the parent-company test is significant. It shifts enforcement from border control toward compute governance: who owns the entity, who controls the data center, who rents the cluster, and who benefits from the training run.

Nvidia’s own FY2026 10-K shows how broad the compliance perimeter already is. The company disclosed licensing requirements affecting products above performance thresholds, including A100, H100, H800, L40S, RTX 4090, GB200 NVL72, and B200-class products, and noted restrictions tied to parties headquartered in, or ultimately parented from, Country Group D:5 countries including China. It also disclosed a $4.5 billion charge tied to H20 inventory and purchase obligations after April 2025 licensing requirements.

Leakage is real, but not the whole story

A lazy reading says chips will always get through. Some will. The open questions are scale, delay, cost, legal risk, and visibility. Export controls do not need perfect enforcement to matter if they make large AI-training clusters harder to assemble, insure, finance, host, or service.

Malaysia shows why this is not just a China-U.S. border story. It is a legitimate semiconductor and logistics hub and a growing AI data-center market. In July 2025, Malaysia’s trade ministry required Strategic Trade Permits for export, transshipment, and transit of high-performance U.S.-origin AI chips under its Strategic Trade Act framework. That is what happens when a neutral commercial hub becomes part of the enforcement map.

DOJ cases also show why third-country routing is not theoretical. In August 2025, prosecutors alleged AI GPUs were routed through Singapore and Malaysia to conceal shipments to China. In another case, DOJ alleged at least $160 million of Nvidia H100 and H200 GPUs were exported or attempted to be exported in violation of U.S. export laws.

There is still uncertainty. Reuters reported one industry-source estimate that chips moved through the opening could number in the hundreds of thousands, but that is not an official BIS count. The guidance also does not prove every third-country sale to a Chinese-linked entity violated export law. Exposure depends on item classification, timing, license status, knowledge, and end-user facts.

Why it matters

The immediate burden falls on chipmakers, server vendors, cloud providers, data-center landlords, freight forwarders, and governments in hubs such as Malaysia and Singapore. They now need stronger ultimate-parent diligence, end-use checks, server-location audits, and cloud-access controls.

The larger policy bet is clear: frontier AI capability is bottlenecked by scarce controlled compute. If Washington can govern access to that compute across subsidiaries and offshore infrastructure, it can affect the scale and timing of Chinese AI training. If it cannot, the market will route around the rules through more opaque procurement channels. Either outcome is more concrete than the usual tech-cold-war framing.